Thanks for sharing this! Yes, it’s really interesting the advances we’re seeing in AI. I’d also be keen in hearing @john.tavis comment on this new version of Alphafold. I’d also like to link to your conversation with John about this in a previous thread for other users: The Hepatitis B polymerase: A difficult problem begins to crack
Thanks for highlighting this for the community. AlphaFold is a once-in-a-generstion breakthrough that is having enormous impact on the biological sciences. I suspect the authors of it will eventually receive the Nobel prize as they’ve solve a massive problem that resisted cracking for 70+ years. I’m using it essentially daily in my work, where it is guiding both drug discovery and engineering of experimentally useful versions of the HBV polymerase.
The newest version (3) just came out and is even more powerful than before. It has already revealed in my lab the structure of how the HBV RNA initially binds to the polymerase. Unfortunately, they changed the use license on this version, restricting its use for drug discovery. Apparently they have set up thie own drug discovery company and don’t want competition.
Regardless, my postdoc De Razia Tajwar is having a great time with it learning the powerful new things we can do with it.
It’s an open source attempt at reimplementing AlphaFold 3 and it has a decent chance at succeeding. It’s new so there’s no code yet but apparently the author has the reputation of a legend so hopefully it should generate some good results
Also one of the AlphaFold 3 members tweeted:
“Really cool to see the huge engagement surrounding AlphaFold 3 and the structures scientists are posting. We’re working to release the AF3 model (incl. weights) in the next 6 months for academic use, so it won’t depend on our research infra. Also the AFServer job limit is now doubling to 20 per day.”
OpenFold is an open source reimplementation of AlphaFold 2. It’s far more efficient and it gets around the same results
Speaking of someone who’s worked on AI stuff myself (starting in 2016) I can say that looks very optimistic. Particularly that they are able to train it on such tiny datasets and still get very accurate results
As far as AI goes, protein folding is an obvious target with low hanging fruit. It’s because there is a ton of training data out there (known protein structures) and consistently predictable results (as long as you don’t get into the whole thing about proteins changing shapes sometimes, which is a whole other kettle of fish and much harder to address). To oversimplify it, you can almost just take the SMILES codes on one side and the known protein structures on the other and use that is your training set, which is a very convenient clean dataset for an AI researcher. The dataset is usually the hardest part about training a model, just obtaining it and cleaning it etc. Lots of manual labor for most tasks such as making a model that is good with generating text. But that’s not so much the case here, which is great
So I’m optimistic that the results of all this is going to be increasingly open source and producing public domain knowledge that is open to all, even in the next 6 months I expect some developments and breakthroughs
I think this might be addressing the AlphaFold server. @bob I think your link is going to a corporate portal, rather than the public site: https://golgi.sandbox.google.com/about
Does this look like it would be useful to anyone here?
It’s a service where you can design proteins with the help of AI and then have them tested in the real world and just download the results
It sounds like this company is doing a lot of what labs like @john.tavis’ and other molecular @ScienceExperts would do on a daily basis. It’s like outsourcing a standard technique.
Think about it like sending your car to the mechanic to get your headlights changed: you could do it yourself pretty simply if you had the tools and knowhow, but outsourcing it might be easier if you didn’t, or just lacked the time or capacity.